Вы находитесь на странице: 1из 7

IJCSI International Journal of Computer Science Issues, Vol.

10, Issue 4, No 1, July 2013


ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 180

Blind Fake Image detection


Nidhal El Abbadi1, Adil Mohamad Hassan2, Mudher Mohammed AL-Nwany3
1
Computer Science Dep., University of Kufa
Najaf, Iraq

2
Mathematical Dep., University of Kufa
Najaf. Iraq

3
Mathematical Dep., University of Kufa
Najaf. Iraq

Abstract
With the great convenience of computer graphics and However, malicious modification of image content
digital imaging, it becomes much easier to alter the forms a serious threat to the secure and legal usage of
content of images than before without any visually digital images.
traces to catch these manipulations. Many fake By skillful manipulation, forgery may be very
images are produced whose content is feigned. Thus, difficult to recognize by the naked eye. Therefore,
the images cannot be judged whether they are real or automatic detection of image forgery has attracted
not visually. In order to detect fake images, this paper much research interest. In recent years, many image
proposes a blind detection uses singular value forgery detection techniques have been proposed,
decomposition (SVD) as a classifier to make a binary especially passive approaches which do not require
decision on whether an image is fake or real. This any additional information besides the image itself
work is an improvements process to an existence [5] [6].
method to detect fake image using SVD. The Some published methods make use of lighting
experimental results prove the effectiveness of this abnormality [1], blur moment invariants, and
algorithm to detect any small changes in image even similarity/dissimilarity of color and structural
with one dot of real image. characteristics [3].
Digital tools have enabled easy image creation,
Key words: SVD, fake image, singular value modification and distribution, which make fraudulent
decomposition, image processing. image forgeries easier than ever.
Fakes are created either by merging two or more
photos or altering an existing image. Because image
1. Introduction manipulation happens at the pixel level, detection is
not as easy as it was before the digital era. Tricky
Cameras are regarded as trustworthy devices and fakes can be exposed by algorithms that detect
photos traditionally imply truth. Nowadays, digital discrepancies or statistical irregularities at the bit
photos have been widely used as historical records level.
and as evidences of real happenings in applications An image is authentic if it represents a witness to an
from journalist reporting, police investigation, law actual event, place, or time.
enforcement, insurance, medical and dental
A definition of image authenticity should enable us to
examination, military, and museum to consumer
distinguish an authentic image from the fake images,
photography.
such as the 2D composite images and the 3D
While digital photos are conveniently used, their
computer graphics images.
credibility has been severely challenged due to
It is still a problem how to detect whether digital
numerous fraudulent cases involving image forgeries,
images are fake or real. Generally, there is an obvious
e.g. the fake results on human stem-cell research [5].
boundary between the fake area and the real area,
with the improvement of desktop photograph
With the availability of powerful image editing tools, manipulation software, which cannot be used to
numerous image retouching techniques have become distinguish fake images and real images. There are a
practical, which can be used to create great artistic many studies related to detect the fake images. In [5],
works.
a blind detection of photomontage is introduced using
higher order statistics, where photomontage is a

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 4, No 1, July 2013
ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 181

similar concept with image fakery. In [6], a model 2.2 Theorem:


based on bipolar signal perturbation is introduced to
detect spliced images. These two papers used a
, and it's unique.
statistics model and bi-coherence features to detect
image forgery, and they are often used to detect
human speech signal. In [1], Popescu and Farid Proof:
introduced some techniques of exposing digital
forgeries by detecting traces of resampling, which (since and are
also tried to resolve the similar problem. Mahdian
and Saic [2] used periodicity due to interpolation to orthonormal ),
perform blind image authentication. They introduced
Radon transform on the basis of second derivative to Then
detect rotation without estimation of the rotation
angle.
In this paper, a SVD based fake image detection To prove it's unique
scheme is developed, which uses the change of the
direction of the eigenvector in orthogonal subspace to Since is unique (as properties of ),
detect the evidence of image fakery.
And is unique (since )
2. Singular Value Decomposition

In linear algebra, the singular value decomposition So is unique.


(SVD) is a factorization of a real or complex matrix,
with many useful applications in signal processing 2.3 Corollary:
and statistics.
when
SVD is based on a theorem from linear algebra which
says that a rectangular matrix A can be broken down Proof:
into the product of three matrices - an orthogonal
matrix U, a diagonal matrix S, and the transpose of an By theorem 2.2 we have
orthogonal matrix V . The theorem is usually
presented something like this:

=
Then when we have
2.1 Proposition:

If and

then when . when .


Proof:
2.4 Corollary:

If and only if

Then when

And this implies to


Proof:
since
By theorem 2.2

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 4, No 1, July 2013
ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 182

there is specific threshold counted by the following


) is unique and,
relation:

) is unique, Threshold …….. 5

3.1 Fake Image Detection


If and only if
When we have an image and need to check
whether it’s fake or not, it will be decompose using
SVD as follow:
3. Methodology
…….. 6
The proposed method to detect the fake image can
Then the vector is extract from the diagonal
achieve by processing image in many steps as follow:
elements of . The proposed fake image detection
1. First the original image is transformed
can be given as follow:
using SVD:
……………….. 2 .
Where and are the orthogonal matrices,
denotes the transpose of , and is a diagonal Where is the detection value, which denotes the
matrix whose diagonal elements can form a column fake factor of the test image.
vector . if , then the tested image is fake,
2. Two secret column vectors , are otherwise it isn’t fake by Corollary 2.5, In the
constructed, which satisfy current proposed scheme, the two secret vectors,
· || , and , are the key construction, on which the
and || · || detection result depends.
Where denotes the inner product. We supposed
3. The main goal is to protect the image before ‖ ‖=‖ ‖ and ^= ……….… 8
publishing it to the public; this will be achieved by However, the vector ^ is composed of two vectors,
changing the diagonal of matrix result from one is the original vector , and the other is the fake
relation (2) with new elements counted by the vector .
following equation: Relation (7) can rewrite by using the equivalents’ in
( ) relations (8) to get:
. ………… 9
Where is a scalar factor, which is set to 0.0001 for
the purpose of this research.
Relation (9) can be computed approximately as
The vector from relation (3) is restored as new
diagonal elements into zero matrix follow:
correspondingly, for that new image will be
constructed (A') as a protected image from the ║ ║ …… 10
following relation:
by Corollary2.4
…………….. 4
' is the preprocessed image and publish to public.
It is obvious that the detection factor depends on
SVD is robust to slight alteration of images, i.e., the
vector v is stable under slight alteration of the image. the fake vector where and since
In proposed image preprocessing procedure, the are unique (properties of ) so that is
alteration of the vector in relation (2) is very small, unique, while denotes the fake vector of the tested
keep by Proposition 2.2 ,so the image
preprocessing does not change the quality of origin image, so the detection value can reflect the status
image significantly (which will be demonstrated in of fake image. Also, there are two secret vectors in
the later examples). our proposed process, and , which do not have
4. Another suggestion in current research is influence on the absolute value of detection result.
using auto threshold by Theorem 2.3. instead
of constant threshold (0.01) for all images as in
previous researches, which mean for each image

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 4, No 1, July 2013
ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 183

4. The result
Then the counted fake factor of image in fig 2 was
To prove the proposed method we will take some of (1.3462e-004), so we decided by Corollary2.5. this
image and make intended change on it to see how image is original (P < Tth ).
this algorithm works.
4.3 Third example to test small change in copy
4.1 The first image is the famous image (Lenna) of IC3 certificate, the change is made by changing
as shown in fig 1, the origin image will protected by
one number only in the date (1/7/2012 change to
applying relations (3, and 4), it is clear the origin
image have no perceptual difference from the 1/7/2013). Fig 3 shows the two IC3 images.
protected image. Intended changing made on the
protected image by changing the face of Lanna. The
auto threshold for this image which is (1.3555e-004)
counted by relation 5, also the fake factor counting
using relation (10) which is equal (0.0185), it is clear
(P > Tth), then the tested image (C) is fake.

A B

C
Fig 1: (A) origin image. (B) Protected image (image 1). (C) Fake
image

4.2 The second example is to test the effect of


rotation on the result of this method. Protected image
of Lenna (image B in fig 1) is rotated with 90 degree.

Figure 2: rotated image for test (image2)

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 4, No 1, July 2013
ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 184

Figure 3: (A) the protected image (image3). (B) The fake image.

The of this image (2.4055e-004) greater than


threshold of this image (1.0191e-004). So it's fake
by Corollary2.5. .

4.4 The fourth example done on the copy of


questions (Fuzzy mathematics exam) which we
change just one dot on it as shows in fig 4. The
threshold of the protected image was (1.4142e-004)
while the fake factor 1.9060e-004, which is
greater than threshold of this image. So the image is
fake by Corollary2.5.

Fig 4: (A) the protected image (image 4). (B) Fake image.

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 4, No 1, July 2013
ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 185

4.5 To summarize the results from the above These improvements enhance the detection efficiency
examples, and to prove the results, table 1 compares and eliminate false detection, in [7] the false positive
the results from the proposed method with the result rate was 0.8% when checking 1000 image while in
this paper the rate decreases to 0.0%.
by the paper of [7].

Table 1: Results of experiments compared with Lai Chung method We must note that all method of detect fake image
[7]. need the original image and the fake image to make
recognition between them but in current method the
origin image no longer needed.
A comparison between previous work and this work
Method

No. of
image

1 2 3 4 is studied where the sample test taken as color


images.
We can make a decision that our SVD scheme is very
0.01 excellent in detecting fake image and it’s sensitive
Previous method

for any small area modified in any image.

References
PSNR

Average = 38 [1] A.C. Popescu and H. Farid, “Exposing digital


forgeries by detecting
traces of resampling”, IEEE Trans. Signal
Process., vol.53, no.2,
Fake 1.3462e 2.4055 1.9060
0.0185 pp.758–767, Feb. 2005.
factor (P) -004 e-004 e-004
[2] B. Mahdian and S. Saic, “Blind authentication
Proposed method

1.3555e 1.3555e 1.0191 1.4142 using periodic properties


-004 -004 e-004 e-004 of interpolation”, IEEE Trans. Inf. Forensics
Security, vol. 3, no. 3, pp.
107.087 107.087 91.392 87.644 529–538, Sep. 2008.
PSNR
4 4 3 2
[3] Steve J. Leon; “Linear Algebra with
Decision Fake Origin Fake Fake Applications”, Macmillan Publishing
Company, New York; 1996.
Reality
of Fake Origin Fake Fake [4] T. Konda, Y. Nakamura, “A new algorithm for
Images
singular allude composition
and its parallelization”, Parallel Comput.
(2009), doi:10.1016/j.parco
5. Conclusions .2009.02.001

In this paper, we used SVD technique for fake image [5] T.T. Ng, S.F. Chang, and Q. Sun, “Blind
detection Scheme. Before the images are published to detection of photomontage
public, some assistant information is inserted into using higher order statistics”, IEEE Int. Symp.
them. Recurring to the secret information, the work Circuits Syst. (ISCAS),
in [7] introduced the mathematical SVD operation in pp.688–691, May 2004.
fake image.
In this paper a modification of pervious work [6] T.T. Ng and S.F. Chang, “A model for image
especially in [7] is achieved. splicing” ,IEEE Int. Conf.
We calculated the two secret vectors in new method. Image Process. (ICIP), pp.1169–1172, Oct.
The important improvement in this work is the scalar 2004.
factor expanding from 0.01 to 0.0001.
Also there is another improvement related to [7] Wei LU, Fu-Lai Chung, and Hongtao LU,
threshold, where the threshold in [7] was constant for “Blind Fake Image Detection
all images, while in this work the threshold will be Scheme Using SVD”, IEICE TRANS.
different for each image, auto threshold determine for COMMUN., VOL.E89–B, NO.5 2006.
each image instead of constant threshold (0.01).

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 4, No 1, July 2013
ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784
www.IJCSI.org 186

Nidhal El Abbadi, received


BSc in Chemical
Engineering, MSc, and PhD
in computer science,
worked in industry and
many universities, he is
general secretary of
colleges of computing and
informatics society in Iraq,
Member of Editorial board of Journal of Computing
and Applications, reviewer for a number of
international journals, has many published papers and
three published books (Programming with Pascal,
C++ from beginning to OOP, Data structures in
simple language), his research interests are in image
processing, biomedical, and steganography, He’s
Associate Professor in Computer Science in the
University of Kufa – Najaf, IRAQ.

In 1963 born in Najaf city,


Iraq .Has MSc in applied
mathematics from university
of Technology, Baghdad. Has
PhD in fractal geometry. He is
a viewer of scientific journals
and conferences. Member of
the ministerial committee for
updating the career. Member
of ministerial virtual learning committee. More than
25 papers were published in locally journals and
conference.

He graduate from College of


education department of
mathematics from AL-
Mustanseria University/
Baghdad 1994, Worked as a
teacher of Mathematics in
the secondary schools,
currently MSc. student in
university of Kufa, mathematical department.

Copyright (c) 2013 International Journal of Computer Science Issues. All Rights Reserved.